March 20, 2015 17:1 IJAIT S0218213015400102 page 1 1st Reading International Journal on Artificial Intelligence Tools Vol. 24, No. 2 (2015) 1540010 (36 pages) c World Scientific Publishing Company DOI: 10.1142/S0218213015400102 Automatic Extraction of Semantic Relations from Wikipedia ∗ † Patrick Arnold and Erhard Rahm Department of Computer Science, Leipzig University, Augustusplatz 10 Leipzig, 04109, Germany ∗[email protected] †[email protected] Received 12 September 2014 Accepted 22 December 2014 Published We introduce a novel approach to extract semantic relations (e.g., is-a and part-of relations) from Wikipedia articles. These relations are used to build up a large and up-to- date thesaurus providing background knowledge for tasks such as determining semantic ontology mappings. Our automatic approach uses a comprehensive set of semantic pat- terns, finite state machines and NLP techniques to extract millions of relations between concepts. An evaluation for different domains shows the high quality and effectiveness of the proposed approach. We also illustrate the value of the newly found relations for improving existing ontology mappings. Keywords: Information extraction; semantic relations; natural language processing; background knowledge; thesaurus; Wikipedia. 1. Introduction Background knowledge plays an important part in information integration, espe- cially in ontology matching and mapping, aiming at finding semantic correspon- dences between concepts of related ontologies. There are numerous tools and ap- proaches for matching ontologies that mostly focus on finding pairs of semantically equivalent concepts.29,5,28,9 Most approaches apply a combination of techniques to determine the lexical and structural similarity of ontology concepts or to consider the similarity of associated instance data. The lexical or string similarity of concept names is usually the most important criterion. Unfortunately, in many cases the lexical similarity of concept names does not correlate with the semantic concept similarity due to uncoordinated ontology development and the high complexity of language. For example, the concept pair (car, automobile) is semantically matching but has no lexical similarity, while there is the opposite situation for the pair (table, stable). Hence, background knowledge sources such as synonym tables, thesauri and dictionaries are frequently used and vital for ontology matching. 1540010-1 March 20, 2015 17:1 IJAIT S0218213015400102 page 2 1st Reading P. Arnold & E. Rahm The dependency on background knowledge is even higher for semantic ontology matching where the goal is to identify not only pairs of equivalent ontology concepts, but all related concepts together with their semantic relation type, such as is- a or part-of. Determining semantic relations obviously results in more expressive mappings that are an important prerequisite for advanced mapping tasks such as ontology merging30,31 or to deal with ontology evolution.19,15 Table 1 lists the main kinds of semantic relations together with examples and the corresponding linguistic constructs. The sample concept names show no lexical similarity so that identifying the semantic relation type has to rely on background knowledge such as thesauri. Table 1. Semantic concept relations. Relation Type Example Linguistic Relation equal river, stream Synonyms is-a car, vehicle Hyponyms has-a body, leg Holonyms part-of roof, building Meronyms Relatively few tools are able to determine semantic ontology mappings, e.g., S-Match,14 TaxoMap,18 ASMOV22 and AROMA,8 as well as our own approach.2 All these tools depend on background knowledge and currently use WordNet as the main resource. Our approach2 uses a conventional match result and determines the semantic relation type of correspondences in a separate enrichment step. We determine the semantic relation type with the help of linguistic strategies (e.g., for compounds such as “personal computer” is-a “computer”) as well as background knowledge from the repositories WordNet (English language), OpenThesaurus (German language) and parts of the UMLS (medical domain). Together with the match tool COMA23 for determining the initial mapping, we could achieve mostly good results in determining the semantic relation type of correspondences. Still, in some mapping scenarios recall was limited since the available repositories, in- cluding WordNet, did not cover the respective concepts. Based on the previous evaluation results, we see a strong need to complement existing thesauri and dic- tionaries by more comprehensive repositories for concepts of different domains with their semantic relations. To build up such a repository automatically, we aim at extracting semantic correspondences from Wikipedia which is the most comprehensive and up-to-date knowledge resource today. It contains almost any common noun of the English language, and thus presumably most concept names. Articles are user-generated and thus of very good quality in general. Furthermore, Wikipedia content can be accessed free of charge. The rationale behind our approach is based on the observation that definitions in dictionaries or encyclopedias have quite a regular structure. In its classic form, a concept C is defined by a hypernym C, together with some attributes describing 1540010-2 March 18, 2015 10:55 IJAIT S0218213015400102 page 3 1st Reading Automatic Extraction of Semantic Relations from Wikipedia the differences between C and C. As an example, consider the following Wikipedia definition of bicycle: A bicycle, often called a bike, is a human-powered, pedal-driven, single-track vehicle, having two wheels attached to a frame, one behind the other. This definition provides (a) the hypernym of bike, which is a vehicle,and(b)sev- eral attributes to distinguish a bike from the more general concept vehicle. While some attributes like human-powered or pedal-driven are not relevant for ontology mapping, some attributes express part-of relations that are indeed valuable. The phrase having two wheels attached to a frame, for instance, expresses that a bike has wheels and a frame (wheels part-of bike, frame part-of bike). Therefore, defini- tion sentences can provide both is-a and part-of (or its complementary type has-a) relations. Additionally, the definition above provides a synonym relation, as the terms bicycle and bike are obviously equivalent because of the expression “often called ”. From a single definition, we can thus extract three relations of different types: equal, is-a, part-of/has-a. In our work we will show how we can discover the mentioned relations in Wikipedia definition sentence and how we extract the words that take part in such a relation, e.g. {bike, bicycle} is-a {single-track vehicle}. In particular, we make the following contributions: • We present a novel approach to extract semantic concept correspondences from Wikipedia articles. We propose the use of finite state machines (FSM) to parse Wikipedia definitions and extract the relevant concepts. • We use a comprehensive set of semantic patterns to identify all kinds of semantic relations listed in Table 1. The proposed approach is highly flexible and extensi- ble. It can also extract multiple relations from a single Wikipedia article. • We show how we can distinguish between entitiy articles and concept articles by using the categories in which articles are listed. • We evaluate our approach against different subsets of Wikipedia covering different domains. The results show the high effectiveness of the proposed approach to determine semantic concept relations. • We provide a theoretic evaluation on an existing mapping, showing new corre- spondences that can be resolved by the knowledge gathered from Wikipedia. In the next section we discuss related work. Section 3 introduces the notion of semantic patterns and outlines which kinds of patterns we use for discovering se- mantic relations. Section 4 describes the new approach to extract semantic relations from Wikipedia in detail. In Section 5 we evaluate the approach for different test cases from different domains. Finally, we briefly report on applying our approach to the entire Wikipedia and on the use of the new relations for improving existing ontology mappings (Section 6) before we conclude with a summary and outlook (Section 7). 1540010-3 March 18, 2015 10:55 IJAIT S0218213015400102 page 4 1st Reading P. Arnold & E. Rahm 2. Related Work Overcoming the large gap between the formal representation of real-world objects (resp. concepts) and their actual meaning is still an open problem in computer sci- ence. Lexicographic strategies, structured-based strategies and instance data anal- ysis were successfully implemented in various matching tools, but in many mapping scenarios these strategies do not suffice and state-of-the-art tools can neither de- termine a complete mapping, nor can they prevent false correspondences. For this reason, background knowledge sources are highly important, as they can improve the mapping quality where generic strategies reach their limits. Hence, a large amount of research has been dedicated to making background knowledge available in diverse resources. Aleksovski et al. analyzed the value of background knowledge for ontology mapping in detail.1 In particular, they showed that a background on- tology can significantly improve match quality for mapping rather flat taxonomies without much lexicographic overlap. The previous approaches for determining background knowledge and the re- sulting background resources